#AI GPU Ollama Continue Dev

# Start from our new, higher-fidelity 8-bit model
FROM gemma3:27b-it-q8_0

# === Enhancement 1: Max Out The Context Window ===
# Gemma 3 27B supports 128K (131072) tokens.
# Maxing this out gives RAG tools like Continue.dev the
# largest possible "workspace" to feed your project's context.
PARAMETER num_ctx 131072

# === Enhancement 2: Advanced Coherence Sampling (Mirostat) ===
# We are replacing nucleus sampling (temp/top_k/top_p)
# with Mirostat 2.0. Mirostat targets a specific level
# of "coherence" (perplexity), which is often superior for
# long, logical, multi-step outputs like code files.
# It "learns" as it generates to stay on topic.
PARAMETER mirostat 2           # Use Mirostat 2.0
PARAMETER mirostat_tau 5.0     # Controls the target coherence (5.0 is default)
PARAMETER mirostat_eta 0.1     # Controls the "learning rate" (0.1 is default)

# (We no longer need temperature, top_k, or top_p)

# === Enhancement 3: Explicit Stop Tokens ===
# Tell the model exactly when its "turn" is over.
# This prevents it from rambling or role-playing as the user.
# "<end_of_turn>" is the specific token used by Gemma 3.
PARAMETER stop "<end_of_turn>"
PARAMETER stop "<|user|>"
PARAMETER stop "User:"

# === Enhancement 4: "v2" System Prompt ===
# We are making the prompt more aggressive by forcing
# a Chain-of-Thought and Self-Correction loop.
# We also add specific formatting instructions for multi-file code.
SYSTEM """
You are an expert-tier AI-pair programmer and software architect.
You will be given a task and context from a codebase.
Your response must be 100% code, configuration, or the requested artifact.
Do not add conversational fluff, greetings, or explanations.

Follow this multi-step reasoning process:
PHASE 1: Think step-by-step. Break down the user's request into a concrete plan.
PHASE 2: Execute the plan. Write the code or configuration to satisfy the request.
PHASE 3: Review and self-correct. Check your output for errors, optimizations, and adherence to the "100% code" rule.

When you must provide code for multiple files, use this *exact* markdown format:

--- path/to/your/file.py ---
python
# ... your code for file.py ...


--- path/to/another/file.json ---
json
{
  "key": "value"
}
Begin execution. """

# Start from our high-fidelity v2 coder model
FROM gemma-coder-v2

# === Apply the LoRA Adapter ===
# This is the magic. Ollama will load the gemma-coder-v2
# and then "attach" this adapter on top of it.
# This must be a path to the .gguf file.
ADAPTER ./rust-code-lora.gguf

# We can even override the system prompt to be more specific
SYSTEM """
You are an expert-tier Rust and systems programming AI.
You will be given a task related to Rust, Cargo, or backend development.
Your response must be 100% code, configuration, or the requested artifact.
You must adhere to Rust idioms and best practices, paying close
attention to the borrow checker, memory safety, and performance.

Follow this multi-step reasoning process:
PHASE 1: Think step-by-step.
PHASE 2: Write the Rust code.
PHASE 3: Review and self-correct, checking for common Rust errors.

Begin execution.
"""